Approximately half of the survey participants (53%) were female and the mean age of the sample was 48.1 years (SD = 13.6). Education varied from primary or lower secondary education (37.3%), upper secondary or post-secondary non-tertiary education (32.8%), to first or second stage of tertiary qualification (29.9%). The participant BMI values ranged from 17.0 to 40.4, with a mean of 25.2 kg/m2 (SD = 4.0), and interviewer BMI ranged from 18.1 to 39.6, with a mean of 25.3 kg/m2 (SD = 3.9). Cronbach’s alpha for the DEBQ-R items was estimated to be 0.93 in the face-to-face survey and 0.94 in the postal survey.
A two-level regression model was estimated that entered the participant-level variables female gender, age, education, and BMI, and the interviewer-level variable interviewer BMI as fixed effects. The results are reported in the left-most column of Table 1.
The table reveals that the participant characteristics all had a significant positive effect on the DEBQ-R restrained eating scores. Most important to our study is the finding that adjusted for the participant variables, interviewer BMI was positively associated with variations in DEBQ-R in the face-to-face survey. This implies that participants were more likely to report restrained eating behaviors to obese interviewers than to underweight and normal weight interviewers.
It is hard to determine from this result who of the interviewers—underweight or obese—elicited more valid responses to the DEBQ-R, as there is no gold standard available to validate them. To evaluate this issue in the best possible way, we opted for a twofold analysis approach. First, the effect of interviewer BMI was examined for three partially overlapping subgroups, that is, participants interviewed by (1) underweight or normal weight interviewers; (2) normal or pre-obese interviewers; and (iii) pre-obese or obese interviewers, using the WHO Expert Consultation (2004) BMI cut-off points. These cut-offs (in kg/m2) are: <18.5 (underweight), 18.5 to <25 (normal weight), 25 to <30 (pre-obese), 30 to <35 (obese class I), 35 to <40 (obese class II), ≥40 (obese class III). The results presented in the second to fourth column of Table 1 reveal that interviewer BMI had a positive effect among participants interviewed by underweight or normal weight interviewers and among those questioned by pre-obese or obese interviewers. Among participants interviewed by normal or pre-obese interviewers, however, the effect turned out to be near zero. This finding suggests that underweight interviewers stimulated under-reporting and that obese interviewers induced over-reporting of dietary restraint.
A second approach was to examine the responses to the postal questionnaire and, in particular, intra-individual differences between the face-to-face and the postal survey responses. The postal follow-up was completed by part of the respondents of the cross-sectional study. Although there were no indications that the postal completers were a biased selection of the baseline responders, we re-ran the two-level regression model for the 504 participants who completed the follow-up. The results, reported in the fifth column of Table 1, indicate that the parameter estimates were largely equivalent to the baseline results. Most notable is that the positive effect of interviewer BMI remained statistically significant. We subsequently analyzed the responses to the DEBQ-R in the postal questionnaire benchmark, to confirm the absence of an interviewer effect in this non-interviewer-administered survey mode. To do so, the postal participants were nested within their former face-to-face interviewers. As can be seen in the sixth column of Table 1, the self-administered postal responses were indeed unaffected by interviewer BMI. To investigate within-participant changes the DEBQ-R difference scores were obtained by subtracting the postal survey scores from the face-to-face interview responses. The estimates, displayed in the right-most column of Table 1, indicate that interviewer BMI is the only characteristic that had a positive effect on the ∆DEBQ-R difference scores. This finding again suggests that answers to questions about dietary restraint are related to the interviewer in that underweight interviewers coaxed lower scores on the DEBQ-R in the face-to-face survey than obese interviewers.
The amount of under- and over-reporting predicted by the regression models is graphically represented in Fig. 1. The effect of interviewer BMI on misreporting of the DEBQ-R in the face-to-face survey (N = 1,212) is displayed by the solid line and scaled on the left-hand side axis. The effect of interviewer BMI on the DEBQ-R difference scores (n = 504) is displayed by the dashed line and scaled on the right-hand side axis of ordinates.
The figure shows that the DEBQ-R scores and the DEBQ-R difference scores yield about the same result as to under- and over-reporting. They both indicate that normal and pre-obese interviewers obtained valid responses to the eating questionnaire in the face-to-face survey, that underweight interviewers stimulated a one-point under-reporting and that obese interviewers triggered a two-point over-reporting in the personal interview. The difference between participants interviewed by underweight and those questioned by obese class II interviewers amounts to approximately three points (i.e., 7.5%) on the DEBQ-R scale.
We additionally performed several sensitivity analyses and supplementary tests. In brief, the results with respect to misreporting obtained by the change score method used here are near equivalent to those obtained using the regression variable method, where the DEBQ-R postal score is regressed on the participant-level variables and interviewer BMI, while controlling for the DEBQ-R face-to-face score (Allison 1990). Also, the models presented in Table 1 include interviewer BMI as a metric variable. Similar results are obtained if it is included as a non-metric grouping variable using the WHO BMI cut-offs. We also examined potential cross-level interactions of the participation characteristics and interviewer BMI to see if interviewer BMI is more influential in some participants. The likelihood ratio tests indicated that the interactions between gender, age and education on the one hand and interviewer BMI on the other are not statistically significant and may be omitted from the regression model without a significant decrease in model fit.